Ray Kurzweil bet big on exotic substrates. In one stretch of The Singularity Is Near, he pointed at gallium-arsenide transistors running at 604 gigahertz, rotaxane memory at 100 gigabits per square inch, IBM polymers self-assembling into 20-nanometer hexagons, and spin currents flowing without energy loss, and told readers these were the off-ramps Moore’s law would take once silicon hit the wall. Twenty years on, the interesting question isn’t whether he was right about the destination — raw compute blew past his targets. The question is how many of his named vehicles actually made the trip. Short answer: one.

The predictions

Batch 5 clusters twelve 2005 forecasts about computing substrates. A few are observations of 2004–2005 lab demos Kurzweil expected to scale. Most project forward: DRAM bits per dollar doubling every 1.5 years; supercomputers matching the human brain “by the end of the 2000s”; a $1,000 personal computer matching one brain “around 2020”; the ITRS roadmap carrying silicon through 2018; and by 2025, nanotube logic switching millions of times faster than neurons.

Read together, these are a bet on a specific story: Moore’s law keeps going because the sixth paradigm — molecular, spin-based, and self-assembling devices — steps in when planar silicon stalls. Kurzweil wrote that “spintronics will play an important role in future computer memory and likely contribute to logic systems, especially in quantum computing” (ch. “Computing with Spin”), and that “rotaxane-based memory and switching devices had been demonstrated with the potential to store one hundred gigabits per square inch” (ch. “Computing with Molecules”). Those names mattered. He wasn’t just saying “compute gets cheaper.” He was saying how.

Where we actually are

Supercomputers vs. the brain. Kurzweil’s conservative 2005 estimate put the brain at 10^16 operations per second. In The Singularity Is Nearer (2024), he revises downward, writing that “total brain computation could be as low as around 10^13 operations per second” and adding that “as of 2023, about $1,000 worth of hardware can already achieve this.” Lawrence Livermore’s El Capitan verified 1.809 exaflops on the November 2025 TOP500; Oak Ridge’s Frontier sustains 1.35 exaflops. Against Kurzweil’s revised brain estimate, El Capitan is roughly ten thousand human brains of raw compute; against his 2005 upper bound, still about 180. Either way, the “end of the 2000s” deadline slipped. Blue Gene/P shipped November 2007 at 167 teraflops at Forschungszentrum Jülich — an order of magnitude below the 10^15 cps Kurzweil quoted for that launch — and didn’t cross 1 petaflop until its 2009 JUGENE upgrade. The brain got matched. It took about a decade longer than advertised, and when it happened the machines were GPUs in liquid-cooled cabinets, not the molecular fabric Kurzweil sketched.

Spintronics: the one prediction that landed clean. The global MRAM market was $912 million in 2025, projected by Future Market Insights to reach $4.77 billion by 2035, with spin-transfer-torque variants expected to hold 52% share. Everspin reported $14.8 million in Q4 2025 revenue, up 12% year-over-year, and Shanghai Siproin Microelectronics began shipping China’s first commercial STT-MRAM chips in 2025, manufactured on SMIC’s 40-nanometer process at densities up to 256 megabits. The patent record is concrete. US 12,408,557, granted September 2, 2025 to Applied Materials, covers methods for forming magnetic tunnel junction film stacks with pinned, tunnel barrier, and free layer crystallinity tuned for STT-MRAM fabrication. US 12,178,137, granted December 2024, describes an in-array magnetic shield — an electrically conductive structure embedded in an interconnect dielectric — protecting MTJ cells from stray fields. US 12,114,578 adds a diffusion-blocking layer of bismuth, antimony, osmium, or rhenium on top of the free layer to keep reference and free layers from contaminating each other during back-end thermal cycles. These are the mundane, high-volume manufacturing patents you file when a memory technology has crossed from curiosity into product. IBM filed eleven STT-MRAM patents since 2020, with TSMC, Samsung, SanDisk, Applied Materials, and Western Digital close behind. The part Kurzweil undersold: spintronic logic has not emerged. This is a memory story, not a computing-with-spin story.

Rotaxane memory. In 2007, Stoddart’s group at UCLA reported a 160-kilobit rotaxane crossbar at 10^11 bits per square centimeter — the density “predicted for commercial memory devices in approximately 2020.” It never shipped. Stoddart won the 2016 Nobel in Chemistry for the broader mechanically-interlocked-molecule field, but rotaxane memory stayed in the lab. A handful of patent filings per year, no large-corporate push. What beat it was boring: vertical NAND stacking went from 24 layers in 2013 to 200-plus in 2024, delivering the density gains without the molecular chemistry. Wrong mechanism, not a missed one.

IBM’s self-assembling polymer memory. The 2003 demo Kurzweil cited used block copolymer lithography. Patent filings in that area ramped through 2014–2016 at about seven per year, then collapsed. EUV lithography absorbed the self-assembly momentum. ASML shipped its first high-NA EUV tool in 2024, and the industry now routes sub-20nm patterning through $400-million photolithography machines, not polymer chemistry.

The 604-gigahertz transistor. In 2025, a group published in Nature Electronics demonstrating metal-oxide-semiconductor field-effect transistors built on aligned carbon nanotube films with a cut-off frequency beyond 1 terahertz at an 80-nanometer gate length — more than 1.6x past the number Kurzweil cited as the frontier. Separately, Northrop Grumman’s DARPA-funded solid-state amplifier demonstrated 9 dB gain at 1.3 THz. Carbon nanotube transistor patent filings peaked around 2017 and have declined as CNT research migrated into high-mobility channel studies and flexible electronics. The THz speed arrived, five years later than implied, in a lab — not in a production fab.

DRAM bits per dollar doubling every 1.5 years. The cleanest miss. AI Impacts’ data shows the historical rate held from roughly 1957 to 2010. After 2010, the doubling time lengthened to about 14 years for a 10x price decline. The 2024–2026 HBM-driven pricing spike made things worse — AI memory demand pushed DDR5 and HBM3E contract prices up through 2025. Compute-per-dollar still bends upward, but because of GPU architecture, not DRAM density.

$1,000 of compute = one human brain by 2020. Kurzweil writes that “as of 2023, $1,000 of computing power could perform up to 130 trillion computations per second.” That’s 1.3 × 10^14 — roughly equal to his revised brain estimate. By his 2024 rubric, he hit the target about three years late. By his original 2005 rubric, which assumed 10^16, a single $1,000 machine still doesn’t get there; an RTX 5090 lands around 1.6 × 10^14 FP16 ops/sec. The verdict depends on which estimate you pin him to.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across this batch is consistent. Kurzweil correctly named the decade in which raw compute would catch up to human-brain-scale processing, and in some framings he was even conservative. What he got wrong was the vehicle. He picked the exotic substrates visible in 2003–2005 lab results, and in every named case other than spintronic memory, something more prosaic won. Block copolymer self-assembly lost to EUV. Rotaxane crossbars lost to vertical NAND stacking. Nanotube logic lost to FinFETs and then gate-all-around. Gallium-arsenide high-frequency transistors stayed in radar and radio.

The surviving named technology, spintronic memory, is instructive because of what it doesn’t do. STT-MRAM replaced neither DRAM nor SRAM in the way Kurzweil’s broader “spin-based computing” framing implied. It sells into niches — automotive, aerospace, IoT microcontrollers, small on-die caches — where non-volatility and radiation tolerance outweigh density. The 2035 forecast is real growth, but it’s a $4.8 billion market in a semiconductor industry approaching a trillion. His directional bet was right. His breadth of claim was not.

The meta-lesson: curve-fitting to exponentials is safer than naming which substrate will carry the curve. The exponential is a statement about economic pressure and all the bets placed against it. The substrate is one bet. Kurzweil’s 2024 revisions — cutting his brain estimate by two orders of magnitude while holding compute projections constant — are an implicit admission that the destination was always more durable than the route.

Method note

We scored this batch by pulling full-text patent evidence for each named technology — STT-MRAM, rotaxane memory, self-assembling polymer lithography, and carbon nanotube transistors — from a corpus of 9.3 million US patent grants and applications, and cross-checking against a scientific literature corpus of 357 million works. For industry milestones we drew on Nature Electronics, TOP500, AI Impacts, and Future Market Insights, and quoted Kurzweil’s 2024 follow-up directly to measure him against his current position, not a decades-old one. Verdicts reflect evidence as of April 2026.